Naïve Forecasting Methods

①  Simple Naïve Model

Q  Naive Trend Model

③  Naive Seasonal Model

 Naive Trend and Seasonal Model

⑤  Naive Rate of Change Model

43
Q 適合無趨勢的時間序列，即數據沒有明顯的上升或下降趨勢，呈現⽔平波動。

Simple Naive Model

 Suitable for no trend/ horizontal series
Yt = 0 + t, t ~ N(0, 2)

where 0 may change slowly with time

 The forecast for the time period t+1 at time

period t is equal to the actual value in the
current period t .
Ft 1  Y t
44
 Quarterly Sales

Simple Naïve 800

700

Model Example 600

500

Units
400
300
200

Example: Sales of ovens for 100

0
the ABC company 1 3 5 7 9 11 13 15 17 19 21 23
Quarter

Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550 Forecast
5 550 400 350 600
6 750 500 400 650 # Actual value
F 25  Y 24  650
7 ?
45
 Simple Naive Model
 The forecast for the time period t+2 at time period t is equal to

Ft  2  Yt 1 (unknown)  Ft 1  Yt
 The forecast for the time period t+3 at time period t is equal to

Ft 3  Yt  2 (unknown)  Ft  2  Yt
 In general, the forecast for the time period t+m at time period t
is equal to the actual value in the current period t .

Ft  m  Yt Yt Yt+1 Yt+2 …. Yt+m

46
 Quarterly Sales

Simple Naïve 800

700

Model Example 600

500

Units
400
300
200

Example: Sales of ovens for 100

0
the ABC company 1 3 5 7 9 11 13 15 17 19 21 23
Quarter

Quarter
Year 1 2 3 4 F2 5  Y 2 4  6 5 0
1 500 350 250 400 F2 6  Y 2 4  6 5 0
2 450 350 200 300 F2 7  Y 2 4  6 5 0
3 350 200 150 400 F2 8  Y 2 4  6 5 0
4 550 350 250 550
5 550 400 350 600
6 750 500 400 650
7 ?
47
②
Naïve Forecasting Methods
 Simple Naïve Model
 Naive Trend Model
 Naive Seasonal Model
 Naive Trend and Seasonal Model
 Naive Rate of Change Model

48
 Naïve Trend Model
適合具有線性趨勢的時間序列數據。這意味著數據隨著時間有增長或下降的趨勢。

 Suitable for linear trend series

Yt   0  1t   t ,  t ~ N 0,  2 
 The forecast for the time period t+1 at time period t
is equal to the sum of the actual value in the current
period t and the estimated trend.
-
Ft+1 = Yt + estimated trend = Yt + (Yt – Yt-1)

49
 Naïve Trend Model Example
 Example: What is the forecast demand for period 25 if
the trend effect is considered?
Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 -
400 650
7 ?
F25  Y24  (Y24  Y23 )
 650  (650  400 )
 900
50
 Naive Trend Model
 The forecast for the time period t+2 at time period t is equal to

Ft  2  Yt 1  Yt 1  Yt   Ft 1  Ft 1  Yt 
 Yt  Yt  Yt 1   [Yt  Yt  Yt 1   Yt ]  Yt  2Yt  Yt 1 
 The forecast for the time period t+3 at time period t is equal to

Ft 3  Yt  2  Yt  2  Yt 1   Ft  2  Ft  2  Ft 1 
 Yt  2Yt  Yt 1   [Yt  2Yt  Yt 1   Yt  Yt  Yt 1 ]  Yt  3Yt  Yt 1 

 In general, the forecast for the time period t+m at time period t is
↑

Ft  m  Yt  Yt  Yt 1  * m
51
 Naïve Trend Model Example
Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 #is
400 11
650 Ye4
7
AFES
?
EEO LFIF AFzS
F2 6  F2 4  2  Y 2 4  ( Y 2 4  Y 2 3 ) * 2
 650  (650  400 ) * 2  1150
F2 7  F2 4  3  Y 2 4  ( Y 2 4  Y 2 3 ) * 3
 650  (650  400 ) * 3  1400
F2 8  F2 4  4  Y 2 4  ( Y 2 4  Y 2 3 ) * 4
 650  (650  400 ) * 4  1650
52
 Naïve Forecasting Methods
 Simple Naïve Model
 Naive Trend Model
 Naive Seasonal Model
 Naive Trend and Seasonal Model
 Naive Rate of Change Model

53
 適合於具有固定季節性週期的時間序列數據，例如季度、⽉度或年度的數據。

假設未來的數值等於⼀年前（或⼀個完整季節長度前）同⼀時期的實際觀測值。

Naïve Seasonal Model

 Suitable for seasonal series
Yt = St + t
where the seasonal variation St may change slowly over
time.
 The forecast for the time period t+1 at time period t is
equal to the actual value in the corresponding season one
year ago.
 For quarterly data, seasonal length = 4. The forecast for
the time period t+1 at time period t is equal to the
actual value in the corresponding quarter one year ago.
Ft+1 = Yt+1-4 = Yt-3
 For monthly data, seasonal length = 12. The forecast
for the time period t+1 at time period t is equal to the
actual value in the corresponding month one year ago.
Ft+1 = Yt+1-12 = Yt-11
54
Naïve Seasonal Model Example
What is the forecast demand for period 25
given a seasonal length of 4 (quarters)?
Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 400 650
7 ?
F25  Y254  Y21  750
55
 Naive Seasonal Model
 For quarterly data, the forecast for the time period t+2 , t+3, t+4,
t+5, t+6 at time period t is equal to
Ft  2  Yt  2 4  Yt  2
Ft 3  Yt 3 4  Yt 1
Ft  4  Yt  4 4  Yt
Ft 5  Yt 5 4  Yt 1 (unknown)  Ft 1
Ft  6  Yt  6 4  Yt  2 (unknown)  Ft  2
 In general, the forecast for the time period t+m at time period t is

Ft  m  Yˆt  m  L
Use actual values for Yˆt  m  L when possible;otherwise use forecasts
56
 Naïve Seasonal Model Example
Quarter
Year 1 2 3 4 TIME =

1 500 350 250 400

2 450 350 200 300 -
3 350 200 150 400
4 550 350 250 550
5
6
550
750
400
500
350
400
600
650
ATL
F25  Y254  Y21  750
7 ? F26  Y264  Y22  500
F27  Y274  Y23  400
F28  Y284  Y24  650

57
 Naive Seasonal Model
 For monthly data, the forecast for the time period t+2 , t+3, t+4, … ,
t+13, t+14 at time period t is equal to
Ft  2  Yt  212  Yt 10
Ft 3  Yt 312  Yt 9
Ft  4  Yt  412  Yt 8

Ft 13  Yt 1312  Yt 1 (unknown )  Ft 1
Ft 14  Yt 1412  Yt  2 (unknown )  Ft  2
 In general, the forecast for the time period t+m at time period t is

Ft  m  Yˆt  m L
Use actual values for Yˆt  m  L when possible;otherwise use forecasts
58
 Naïve Forecasting Methods
 Simple Naïve Model
 Naive Trend Model
 Naive Seasonal Model
 Naive Trend and Seasonal Model
 Naive Rate of Change Model

59
 適合⽤於同時具有線性趨勢和加法性季節性的時間序列數據

Naïve Trend and Seasonal Model

 Suitable for linear trend and additive seasonality
series
Yt = (0 + 1t) + St + t

where 0 , 1 , St may change slowly with time and

the seasonal variation is constant over time.

 The forecast for the time period t+1 at time period t

is equal to the sum of the estimated trend and
seasonality at time period t.

60
 Naïve Trend and Seasonal Model
 For quarterly data, the Yt-3 term estimates the seasonal
patterns and the remaining term which averages the
amount of change for the past 4 quarters estimates the
TAGE trend
HERE)
-EGM (Yt  Yt 1 )  (Yt 1  Yt 2 )  (Yt 2  Yt 3 )  (Yt 3  Yt 4 ) Yt  Yt 4
data.Ft 1 It
Yt 3   Yt 3 
4 4

 For monthly data, the Yt-11 term estimates the seasonal

patterns and the remaining term which averages the
amount of change for the past 12 months estimates the
trend
(Yt  Yt 1 )  (Yt 1  Yt 2 )    (Yt 11  Yt 12 ) Y  Yt 12
Ft 1  Yt 11   Yt 11  t
12 12
61
 Naïve Trend and Seasonal Model Example
 Example: To find the forecast demand for period 25
Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 400 650
7 ?
(Y24  Y23)  (Y23  Y22 )  (Y22  Y21)  (Y21  Y20 )
F25  Y21 
4
(Y  Y )
 Y21  24 20
4
(650 600)
 750
4
 762..5
62
 Naive Trend and Seasonal Model
 For quarterly data, the forecast for the time period t+2 , t+3, t+4,
t+5, t+6 at time period t is equal to
Y Y   F Y 
Ft  2  Yt  24   t  21 t  214   Yt 2   t 1 t 3 
 4   4 
Y Y   F Y 
Ft 3  Yt 34   t 31 t 314   Yt 1   t  2 t 2 
 4   4 
 Yt  41  Yt  414   Ft 3  Yt 1 
Ft  4  Yt  44     Yt   
 4   4 
 Yt 51  Yt 514   Ft  4  Yt 
Ft 5  Yt 54     Ft 1   
 4   4 
 Yt 61  Yt 614   Ft 5  Ft 1 
Ft 6  Yt 64     Ft  2   
 4   4  63
 Naive Trend and Seasonal Model
 For monthly data, the forecast for the time period t+2 , t+3, t+4, … ,
t+13, t+14 at time period t is equal to
Y Y   F Y 
Ft  2  Yt  212   t  21 t  2112   Yt 10   t 1 t 11 
 12   12 
 Yt 31  Yt 3112   Ft  2  Yt 10 
Ft 3  Yt 312     Yt 9   
 12   12 
 Yt  41  Yt  4112   Ft 3  Yt 9 
Ft  4  Yt  412     Yt 8   
 12   12 
Y Y   F Y 
Ft 13  Yt 1312   t 131 t 13112   Ft 1   t 12 t 
 12   12 
Y Y   F  Ft 1 
Ft 14  Yt 1412   t 141 t 14112   Ft 2   t 13 
 12   12  64
 Naive Trend and Seasonal Model
 In general, the forecast for the time period t+m at
time period t is

 Yˆ  Yˆ 
Ft  m  Yˆt  m  L   t  m 1 t  m 1 L


 L 
Use actual values for Yˆt m  L , Yˆt  m 1 , Yˆt  m 1 L
when possible; otherwise use forecasts.

65
Naïve Trend and Seasonal Model Example
Quarter
Year 1 2 3 4 &

1 500 350 250 400

2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 400 650
S

7 ?

(Y24  Y20 ) (650  600)

F25  Y21   750   762.5
4 4
( F25  Y21 ) (762.5  750)
F26  Y22   500   503.1
4 4
( F26  Y22 ) (503.1  500)
F27  Y23   400   400.8
4 4
(F  Y ) (400.8  400)
F28  Y24  27 23  650   650.2
4 4
66
 Naïve Forecasting Methods
 Simple Naïve Model
 Naive Trend Model
 Naive Seasonal Model
 Naive Trend and Seasonal Model
 Naive Rate of Change Model

67
 這是⼀種基於變動率的簡單預測模型，適合⽤於數據隨著時間變動率相對穩定的情況。

Naive Rate of Change Model

 For some purposes, the rate of change might be more appropriate
than the absolute amount of change.
 The forecast for the time period t+1 at time period t is equal to
the product of the actual value in time period t and the estimated
rate of change.  Y 
Ft 1  Yt  t 
 Example  Yt 1 
Quarter  Y 24 
Year 1 2 3 4 F 25  Y 24  
1 500 350 250 400  Y 23 
2 450 350 200 300
3 350 200 150 400  650 
4 550 350 250 550  650  
5 550 400 350 600  400 
6 750 500 400 650  1056
7 ?
68
 Naive Rate of Change Model
 The forecast for the time period t+2 at time period t is equal to
  Yt  
 Yt    2
 Yt 1   Ft 1   Yt    Yt 1    Yt 
Ft  2  Yt 1    Ft 1    Yt    Yt  
 Yt   Yt   Yt 1   Yt   Yt 1 
 
 
 The forecast for the time period t+3 at time period t is equal to
  Y 2 
2  Yt 

t
  3
 Yt  2   Ft 2   Yt    Yt 1    Yt 
Ft 3  Yt 2    Ft 2    Yt    Yt  
 
 Yt 1   Ft 1   Yt 1   Y  Yt    Yt 1 
t
Y 
  t 1  
m
 In general, the forecast for the time  Yt 
period t+m at time period t is Ft m  Yt  
 Yt 1  69
 Naive Rate of Change Model
Quarter
Year 1 2 3 4
1 500 350 250 400
2 450 350 200 300
3 350 200 150 400
4 550 350 250 550
5 550 400 350 600
6 750 500 400 650
7 ?
 Y 24   650 
F25  F241  Y 24    650    1056
Y
 23   4 0 0 
2 2
 Y 24   650 
F26  F24 2  Y 24    650    1716
Y
 23   400 
3 3
 Y 24   650 
F27  F24 3  Y 24    650    2789
 Y 23   400 
4 4
 Y 24   650 
8
F28  F24 4  Y 24  
 Y 23 
 650  
 400 
 3 1, 6 0 4
70
 Part Four

 Introduction to Forecasting
 Time Series Components
 Naive Forecasting Methods
 Measuring Forecast accuracy

71
 Measuring Forecast Accuracy
Forecast error is the difference between the actual value
and the predicted value =actual -predicted
e e

0 0

T T
Random errors Cyclical effects not accounted for
e e
0 0

T T
Trend not accounted for Seasonal effects not accounted for
 Measuring Forecast Accuracy
 Forecast error or residual e t  Y t  Ft
n
 Mean error ME   et n
t 1

n
 Mean absolute error MAE   et n
t 1
n
 Mean squared error MSE   et2 n
t 1
73
Example

ME 
 e 15  15.5    23  27 
 i
 1
n 4

MAE 
 ei 15  15.5    23  27
  1.25
n 4

MSE 
i 
e 2
15  15.5 2
   23  27 2
 4.125
n 4 74
 Measuring Forecast Accuracy
 Percentage error PE t  e t Y t   100

n
 Mean percentage error MPE   PEt n
t 1

 Mean absolute percentage error

n
MAPE   PEt n
t 1

75
Example

ei   15  15.5   23  27  
Y * 100   15 
 
   
 23  

MPE  i
  * 100  4.523%
n  4 
 
ei  15  15.5 23  27 
Y * 100
 15   
23 
<10% highly accurate
10-20% good
MAPE  i
  * 100  5.839% 20-50% reasonable
n  4  >50% inaccurate
 
76
 Measuring Forecast Accuracy

 Theil’s U-statistic
 This statistic allows a relative
comparison of formal forecasting
methods with naïve approaches and
also squares the errors involved so that
large errors are given much more
weight than small errors.

77
 Definition of Theil’s U-statistic
n 1 n

{( Ft 1  Yt 1 ) / Yt } 2
 t t t 1
{( F  Y ) / Y }2

U t 1
n 1
 t 2
n

 t 1 t t
{(Y
t 1
 Y ) / Y }2
 t t 1 t 1
{(Y
t 2
 Y ) / Y }2

2 2
 F2  Y2   F3  Y3 
    
 Y1   Y2 
 2 2 U  1 if Ft 1  Yt
 Y2  Y1   Y3  Y2 
    
 Y1   Y2 
78
 Interpreting Theil’s U-statistic
U = 1: The naïve method is as good as the forecasting
technique being evaluated.

U < 1: The forecasting technique being used is better

than the naïve method. The smaller the U-
statistic, the better the forecasting technique is
relative to the naïve method.

U > 1: There is no point in using a formal forecasting

method, since using a naïve method will produce
better results.

79
 n 1 n

{( F t 1  Yt 1 ) / Yt } 2
{( F  Y ) / Y
t t t 1 }2
U t 1
 t 2

Example
n 1 n

{(Y
t 1
t 1  Yt ) / Yt }2 {(Y  Y
t 2
t t 1 ) / Yt 1}2

[(Predictedt+1 - [(Actualt+1 -
Year Actual Sales Forecast Actualt+1)/Actualt]2 Actualt)/Actualt]2
1 15 15.5 0.0000 0.1111
2 20 20 0.0006 0.0025
3 19 18.5 0.0443 0.0443
4 23 27
sum= 0.0449 0.1579
sqrt= 0.2120 0.3974
Theil's U 0.5335

[(Predictedt - [(Actualt -
Year Actual Sales Forecast Actualt)/Actualt-1]2 Actualt-1)/Actualt-1]2
1 15 15.5
2 20 20 0.0000 0.1111
3 19 18.5 0.0006 0.0025
4 23 27 0.0443 0.0443
sum= 0.0449 0.1579
sqrt= 0.2120 0.3974
Theil's U 0.5335
80
Example
t 1
2 2 2
 ( Ft 1  Yt 1 )   ( F2  Y2 )   20  20 
        0
 Yt   Y1   15 
2 2 2
 (Yt 1  Yt )   (Y2  Y1 )   20  15 
         ( 0.333) 2
 0.111
 Yt   Y1   15 

t2
2 2 2
 ( Ft 1  Yt 1 )   ( F3  Y3 )  19  18.5 
    Y    20   0.000625
 Y t   2 
2 2 2
 (Yt 1  Yt )   (Y3  Y2 )  19  20 
    Y    20   0.0025
 Y t   2 
81
Example
t3
2 2 2
 ( Ft 1  Yt 1 )   ( F4  Y4 )   27  23 
    Y    19   0.04432
 Y t   3 
2 2 2
 (Yt 1  Yt )   (Y4  Y3 )   23  19 
    Y    19   0.04432
 Y t   3 

n 1

 {( F t 1  Yt 1 ) / Yt }2

0.04494633
U t 1
n 1
  0.5335
0.157932441
 {(Y
t 1
t 1  Yt ) / Yt }2

82
MS4212 Predictive Analytics and
Forecasting

Chapter 2 Moving Average Methods

1
Part One
 Simple moving average
 Double moving average
 Application of moving average to stock price

3
Simple Moving Average
 Suitable for no trend/ horizontal series
yt = 0 + t, t ~ N(0, 2)
where 0 may change slowly with time
 The forecast for time period t+1 at time period t is equal to

the n-period moving average calculated at time period t,

which is the simple average of the n most recent
observations y  y  y t t 1 t ( n1)
Ft 1  Mt 
n
t-3 t-2 t-1 t t+1 t+2

n observations

4
Simple Moving Average
 As each new observation (eg yt+1 ) becomes available, a new
moving average (eg Mt+1 ) and hence a new forecast (eg Ft+2 ) can
be computed by dropping the oldest observation and adding the
newest observation. The forecast for time period t+2 calculated at
time period t+1 is equal to
yt 1  yt    yt ( n2)
F(t 1)1  Ft 2  M t 1 
n
 In general, the forecast for time period t+m calculated at time
period t+m-1 is equal to
yt  m 1  yt  m  2    yt  m  n
F(t  m 1) 1  Ft  m  M t  m 1 
n
where m = 1, 2, 3, … is the forecast horizon.
 Note: if actual Y values are unavailable, use forecast F values
5
 Example 1 (Level Data): Simple Moving Average

Time Observed 3-month 3-mth MA 5-month 5-mth MA

Month Period (t) Values(Yt) MA (Mt) Forecast(Ft) MA(Mt) Forecast(Ft)
Jan 1 200
Feb 2 135
Mar 3 195 176.7 FAL/
Apr 4 197.5 175.8 176.7
May 5 310 234.2 175.8 207.5
Jun 6 175 227.5 234.2 202.5 207.5
Jul 7 155 213.3 227.5 206.5 202.5
Aug 8 130 153.3 213.3 193.5 206.5
Sep 9 220 168.3 153.3 198.0 193.5
Oct 10 277 209.0 168.3 191.4 198.0
Nov 11 235 244.0 209.0 203.4 191.4
Dec 12 244.0 203.4
Example 1 (Level Data): Simple Moving Average

7
Choice of n for Simple Moving Average
 If n = 1, the simple moving average is simply
the naïve model.
 If n is small, the forecast is more responsive to
fluctuations in data. So use small n if data show
pattern.
 If n is large, the forecast is less responsive to
fluctuations in data. So use large n if data show
large randomness.
 In general, we choose n such that MSE,
MAPE, … is minimum

8
 Time Observed 4-Month 4-Month MA
Period Values MA Forecast
Example 2 (Trend Data): 1 140.00
Simple Moving Average 2 159.00
3 136.00
4 157.00 148.00
5 173.00 156.25 148.00
6 131.00 149.25 156.25
7 177.00 159.50 149.25
8 188.00 167.25 159.50
Simple Moving Average Forecast on 9 154.00 162.50 167.25
Inventory Data 10 179.00 174.50 162.50
300.00 11 180.00 175.25 174.50
250.00 12 160.00 168.25 175.25
13 182.00 175.25 168.25
200.00
14 192.00 178.50 175.25
Inventory

150.00
Observed Values 15 224.00 189.50 178.50
100.00 4‐Month MA Forecast 16 188.00 196.50 189.50
50.00 17 198.00 200.50 196.50
18 206.00 204.00 200.50
0.00
1 3 5 7 9 11 13 15 17 19 21 23 25 19 203.00 198.75 204.00
Time Period 20 238.00 211.25 198.75
21 228.00 218.75 211.25
22 231.00 225.00 218.75
23 221.00 229.50 225.00
24 259.00 234.75 229.50
25 273.00 246.00 234.75
9
26 246.00
Part Two
 Simple moving average
 Double moving average
 Application of moving average to stock price

10
②
Double Moving Average
 Suitable for linear trend series

yt   0  1t   t ,  t ~ N 0,  2 
 Calculation procedures:

 Step 1: The n-period simple moving average at time period t is

calculated by
yt  yt 1    yt  n 1
Mt 
n
 Step 2: The n-period double moving average at time period t is
calculated by
M t  M t 1  M t  2    M t  n 1
M t 
n
11
Double Moving Average
 It can be shown that
 n  1 
E M t   E y t    1
 2 
 n 1
E M t  E M t    1
 Hence,  2 

 n 1 
E ( yt )  E ( M t )  E  M t   E  M t    1
 2 
 2 
1     E  M t   E  M t 
 n 1 
 2 
   E  yt   E  M t  
 n 1 
12
 Tay
Double Moving Average

CIMAL
SIMMACK
Time Actual SMA(3) Act-SMA DMA(3) SMA-DMA A B Forecast eForecast
1 2
2 4
3 6 4 2
4 8 6 2
5 10 8 2 6 2 10 2
6 12 10 2 8 2 12 2 12 0
7 14 12 2 10 2 14 2 14 0
8 16 14 2 12 2 16 2 16 0
9 18 16 2 14 2 18 2 18 0
10· 20 18 2 16 2 20 2 20 0
11 N 22
13
 Double Moving Average

25

20

15 Actual
SMA(3)
10 DMA(3)

0
0 5 10 15
Time
Double Moving Average
 Step 3 : The current level of the data at time t is estimated by

a t  M t   M t  M t  2 M t  M t
 Step 4 : The slope 1 of the series at time t is estimated by
2
bt  M t  M t
n 1

 Step 5 : The forecast for period t + m is obtained by

extrapolating the trend m periods into the future

Ft  m  a t  bt m
15
 𝑌

𝑀
𝑦 ?
𝑀"
𝑦 ?
2
at  M t  M t  M t   2 M t  M t 𝑦 bt  M t  M t
n 1

𝑡 t+1 t+2
 Example 3 (Trend Data) : Double Moving Average
Period Step 1 Step 2 Step 5
Inventory Four-Month Four-Month One-Month
Balance of Moving Moving Ahead
Product Average Average Forecast
E12 of (1) of (2)
1 140.00 -- --- ---
2 159.00 -- --- ---
3 136.00 --- --- ---
4 157.00 148.00 --- ---
5 173.00 156.25 --- ---
6 131.00 149.25 --- ---
7 177.00 159.50 153.25 ---
8 188.00
/ 167.25 158.06 169.92

Step 3 : a7 = 2*M7 –M7’’ = 2*159.5 – 153.25 = 165.75

Step 4 : b7=2*(M7 –M7’’ )/(n-1)=2*(159.5-153.25)/(4-1) = 4.167
Step 5 : F8 = a7 + b7 = 165.75+ 4.167 = 169.917
#
 4-Month
Time Observed Double Value of Value of DMA
Simple
Period Values MA a b Forecast
MA
1 140.00
2 159.00
3 136.00
4 157.00 148.00
5 173.00 156.25
6 131.00 149.25
7 177.00 159.50 153.25 165.75 4.17
8 188.00 167.25 158.06 176.44 6.13 169.92
9 154.00 162.50 159.63 165.38 1.92 182.56
10 179.00 174.50 165.94 183.06 5.71 167.29
Example 3 11 180.00 175.25 169.88 180.63 3.58 188.77
12 160.00 168.25 170.13 166.38 -1.25 184.21
(Trend Data) : 13 182.00 175.25 173.31 177.19 1.29 165.13
14 192.00 178.50 174.31 182.69 2.79 178.48
Double Moving 15 224.00 189.50 177.88 201.13 7.75 185.48
Average 16 188.00 196.50 184.94 208.06 7.71 208.88
17 198.00 200.50 191.25 209.75 6.17 215.77
18 206.00 204.00 197.63 210.38 4.25 215.92
19 203.00 198.75 199.94 197.56 -0.79 214.63
20 238.00 211.25 203.63 218.88 5.08 196.77
21 228.00 218.75 208.19 229.31 7.04 223.96
22 231.00 225.00 213.44 236.56 7.71 236.35
23 221.00 229.50 221.13 237.88 5.58 244.27
24 259.00 234.75 227.00 242.50 5.17 243.46
25 273.00 246.00 233.81 258.19 8.13 247.67
18
26 266.31
 Inventory Forecast
300

250

200
Inventory

150

100

50

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Time Period

Observed values SMA Forecast DMA Forecast

19
 Double Moving Average
y1  y2  y3
M3 
3
t yt Weight
y 2  y3  y 4
M4  1 y1 1/9
3 2 y2 2/9
y3  y 4  y5 3 y3 3/9
M5 
3 4 y4 2/9
5 y5 1/9
M3  M4  M5
M 
"
5
3
y1  2 y2  3 y3  2 y4  y5 1 2 3 2 1
  y1  y2  y3  y4  y5
9 9 9 9 9 9
20
Summary
 Simple moving average
 Equal weight to the past n observations only

 Double moving average

 Unequal weight to a number of the past

observations

21
Part Three
 Simple moving average
 Double moving average
 Application of moving average to stock price

22
Determination of buy/sell
signals
 Use a moving average
 Buy if P > MA
 Sell if P < MA
 Use two moving averages
 Buy if MA(short) > MA (long)
 Sell if MA(short) < MA (long)

24
25
MS4212 PREDICTIVE ANALYTICS AND
FORECASTING

Chapter 3
EXPONENTIAL SMOOTHING METHODS

1
PART ONE
 Simple Exponential Smoothing
 Double Exponential Smoothing
 Brown’s One-parameter Linear Method
 Holt’s Two-parameter Linear Method
 Winters’ Three Parameter Linear and Seasonal
Exponential Smoothing
 Multiplicative Method
 Additive Method

2
SIMPLE EXPONENTIAL SMOOTHING (SES)
 Suitable for no trend series
yt = 0 + t , t ~ N(0, 2)
where 0 may change slowly with time

 The simple exponential smoothed statistic at time t is

At  yt  (1  ) At 1
where  is a smoothing constant between 0 and 1.

 The forecast for period t+1 at time t is

Ft+1 =At = yt + (1  ) At-1
 Another formula for calculating the forecast
Ft+1 = yt + (1  ) Ft = Ft+  (yt – Ft)

 The forecast for period t+m at time period t is

Ft+m = At = yt + (1  ) At-1 , m = 1, 2, 3, 
3
where m is the forecast horizon.
 SIMPLE EXPONENTIAL SMOOTHING
 Expanding the equation by replacing At-1 by its components, At-2 by its components, At3 by its
components and so on, we have

Ft+m = At = yt + (1  )At1

= yt + (1  )[yt1+(1  ) At2]

= yt + (1  )yt1+(1  )2At2

= yt + (1  )yt1+(1  )2[yt2+(1  ) At3]

= yt + (1  )yt1 + (1  )2)yt2 +(1  )3 At3

… = yt + (1  )yt1 + (1  )2 yt2 + (1  )3 yt3

+ (1  )4 yt4 + (1  )5 yt5 + + (1  )t1 y1+
(1  )t A0.

So Ft+m = At represents a weighted moving average of all past observations

where the weights are determined by a smoothing constant.

4
 SES WEIGHTS
• Suppose  = 0.2, 0.4, 0.6 or 0.8. Then the weights assigned to past
observations would be as follows:
time y Weight
value assigned  = 0.2  = 0.4  = 0.6  = 0.8
t yt  0.2 0.4 0.6 0.8

t-1 yt  1 (1-) 0.16 0.24 0.24 0.16

t-2 yt  2 (1-)2 0.128 0.144 0.096 0.032
t-3 yt  3 (1-)3 0.1024 0.0864 0.0384 0.0064
t-4 yt  4 (1-)4 (0.2)(0.8)4 (0.4)(0.6)4 (0.6)(0.4)4 (0.8)(0.2)4

• In each case,
{  (1  )  (1  )2    (1  )t1}  (1  )t
 The weights for all past data sum
approximately to one. [1  (1  )t ]
  (1  )t
 The weights decrease exponentially with 1  (1  )
heavier weights given to recent observations  [1  (1  )t ]  (1  )t  1
and smaller weights given to old 5
observations, hence the name exponential
smoothing.
 SES Weights
0.9

0.8

0.7

0.6

0.5
Weights

0.4

0.3

0.2

0.1

0
t t-1 t-2 t-3 t-4 t-5
Time

Alpha=0.2 Alpha=0.4 Alpha=0.6 Alpha=0.8

 CHOICE OF  IN SES

 The larger the , the more weight given to recent

observations
 If data show large randomness, use small  (usually
0.01 <  < 0.3)
 If data show pattern, use large  (  0.5, suggest
trend or seasonality)
 Choose  which minimize MSE, MAPE,  over a
test set.
7
 INITIALIZATION IN SES

Ft+1 = At = yt + (1  ) At-1

 Use F1 = A0 = y1
 F2 = A1 = y1 + (1  )A0 = y1

 Other alternatives for A1

 A1 = mean of all observations , or
 A1 = mean of the first 4, 5 or 6 observations, or
 A1 = mean of half of the data

8
SIMPLE EXPONENTIAL SMOOTHING
EXAMPLE 1 Ft+1 =At = yt + (1  ) At-1
Forecast
Time Actual At t=1
(α=0.1)
1 200.00 200.00 Step 1: A1 = y1 =200
2 135.00 193.50 200.00 Step 2: F2 = A1 = 200
3 195.00 193.65 193.50 t=2
4 197.50 194.04 193.65 Step 1:
5 310.00 205.63 194.04 A2 =0.1y2 + (1- 0.1) A1
6 175.00 202.57 205.63 = 0.1*135 +0.9*200 = 193.5
7 155.00 197.81 202.57 Step 2: F3 =A2 = 193.5
8 130.00 191.03 197.81
9 220.00 193.93 191.03
10 277.00 202.23 193.93
11 235.00 205.51 202.23 9
12 ? 205.51
SIMPLE EXPONENTIAL SMOOTHING
EXAMPLE 1

10
SIMPLE EXPONENTIAL SMOOTHING
EXAMPLE 2
Forecast t=1
Time Actual At Step 1: A1 = y1 =79
(α=0.1)
1 79.00 79.00 Step 2: F2 = A1 = 79
2 84.00 79.50 79.00
3 83.00 79.85 79.50 t=2
4 81.00 79.97 79.85 Step 1: A2 =0.1y2 + (1- 0.1) A1 = 0.1*84
5 98.00 81.77 79.97 +0.9*79 = 79.5
6 100.00 83.59 81.77 Step 2: F3 =A2 = 79.5
7 83.59

11
SIMPLE
EXPONENTIAL
SMOOTHING
EXAMPLE 2

12
PART TWO
 Simple Exponential Smoothing
 Double Exponential Smoothing
 Brown’s One-parameter Linear Method
 Holt’s Two-parameter Linear Method
 Winters’ Three Parameter Linear and Seasonal
Exponential Smoothing
 Multiplicative Method
 Additive Method

13
 BROWN’S METHOD
 Suitable for linear trend series
Yt = 0 + 1t + t

where 0 and 1 may change slowly with time.

 Calculation procedures for Brown’s method:
 Step 1: The simple exponential smoothed statistic at
time t is A  y  (1  ) A
t t t 1

 Step 2: The double exponential smoothed statistic at

time t is
At"  At  (1  ) At"1
14
BROWN’S METHOD
1 
 It can be shown that E  At   E  yt     1
  
1 
 
E At  E  At   
"
 1
  

1  
 Hence  
E  yt   E  A t   E  At   E At  
"

  
 1

1 
1 

  
E  At   E At  " 
1 
 E  yt   E  At  

15
Time yt(=2t) At yt-At A''t At-A''t at bt Forecast Error BROWN’S
1 2 2.00 0.00 2.00 0.00 2.00 2.00
2
3
4
6
2.80
4.08
1.20
1.92
2.32
3.02
0.48
1.06
3.28
5.14
0.32
0.70
4.00
3.60
0.00
2.40
METHOD
At  yt  (1  ) At 1
4 8 5.65 2.35 4.07 1.57 7.22 1.05 5.84 2.16
5 10 7.39 2.61 5.40 1.99 9.38 1.33 8.27 1.73 At"  At  (1  ) At"1
6 12 9.23 2.77 6.93 2.30 11.53 1.53 10.70 1.30
7 14 11.14 2.86 8.62 2.52 13.66 1.68 13.07 0.93
  0.4
8 16 13.08 2.92 10.40 2.68 15.76 1.79 15.35 0.65
9 18 15.05 2.95 12.26 2.79 17.84 1.86 17.55 0.45 A1  y1  2
10 20 17.03 2.97 14.17 2.86 19.89 1.91 19.70 0.30 A2  0.4* y2  0.6* A1
11 22 19.02 2.98 16.11 2.91 21.93 1.94 21.80 0.20  0.4* 4  0.6* 2
12 24 21.01 2.99 18.07 2.94 23.95 1.96 23.87 0.13
13 26 23.01 2.99 20.04 2.96 25.97 1.97 25.91 0.09
 2.8
14 28 25.00 3.00 22.03 2.98 27.98 1.98 27.94 0.06
15 30 27.00 3.00 24.02 2.98 29.99 1.99 29.96 0.04 A1"  y1  2
16 32 29.00 3.00 26.01 2.99 31.99 1.99 31.98 0.02
17 34 31.00 3.00 28.01 2.99 33.99 2.00 33.98 0.02 A2"  0.4* A2  0.6* A1"
18 36 33.00 3.00 30.00 3.00 36.00 2.00 35.99 0.01  0.4* 2.8  0.6* 2
19 38 35.00 3.00 32.00 3.00 38.00 2.00 37.99 0.01  2.32
20 40 37.00 3.00 34.00 3.00 40.00 2.00 40.00 0.00
21 42 39.00 3.00 36.00 3.00 42.00 2.00 42.00 0.00
22 44 41.00 3.00 38.00 3.00 44.00 2.00 44.00 0.00  1   0.6
 
 1  *2  3
23 46 43.00 3.00 40.00 3.00 46.00 2.00 46.00 0.00    0.4
24 48 45.00 3.00 42.00 3.00 48.00 2.00 48.00 0.00 yt  At  At  A
16
"
t
25 50 47.00 3.00 44.00 3.00 50.00 2.00 50.00 0.00
26 52 49.00 3.00 46.00 3.00 52.00 2.00 52.00 0.00  1  
  1
27 54.00   
17
 BROWN’S METHOD
 Calculation procedures for Brown’s method:
 Step 3: The current level of the data at time t, 0 , is estimated by

at  At   At  At  2 At  At
 Step 4: The slope of the series at time t, 1 , is estimated by
  
bt    At  At
1 
 Step 5: The forecast value for time period t+m made at time t is
Ft+m = at + bt m, m = 1, 2, 3, 
where m is the forecast horizon. 18
 𝑌

𝐴
𝑦 ?
𝐴"
𝑦 ?

bt 
1 
 At  At" 
at  At   At  At"   2At  At" 𝑦

𝑡 t+1 t+2
CHOICE OF  AND INITIALIZATION IN
BROWN’S METHOD A  y  (1  ) A t t t 1

A1  y1  (1  ) A0
 Choice of   y1  (1  ) y1  y1
 Choose one that minimize MSE or MAPE.
At"  At  (1  ) At"1
 Initialization
A1"  A1  (1  ) A0"
 Delurgio A0 = A0 = y1 => A1 = A1 = y1, a1 = y1  y1  (1  ) y1  y1
 b1 = [ (y2 - y1) + (y4 - y3) ] / 2
at  2 At  At
 Or Bowerman (P392)
a1  2 A1  A1
1 
A0  a 0  b0  2 y1  y1  y1
  

 1 
A0  a0  2 b0
  
where a0 and b0 are least squares estimates of 0 and 1 by 20

fitting a straight line to, for example, one half of the data.
 BROWN’S METHOD EXAMPLE
Forecast the demand for period 9 using Brown’s method (=0.4)

Week Actual demand Simple es Double es Forecast

t yt At At” at bt Ft
1 31 31.00 31.00 31.00 9.00
2 40 t=1 40.00
3 43 Step 1: A1 = y1 =31
4 52
Step 2: A ” = y =31
5 49 1 1
6 64 Step 3: a1 = y1 =31
7 58
Step 4: b1 =[ (y2 - y1) + (y4 – y3)]/2
8 68
9 = [(40-31) + (52-43)]/2=9
Step 5: F2 =a1 + b1 = 31 + 9=40 21
 BROWN’S METHOD EXAMPLE
Forecast the demand for period 9 using Brown’s method (=0.4)

Week Actual demand Simple es Double es Forecast

t yt At At” at bt Ft
1 31 31.00 31.00 31.00 9.00
2 40 34.60 32.44 36.76 1.44 40.00
3 43 38.20
4 52 t=2
5 49 Step 1: A2=0.4y2 + 0.6 A1 = 0.4*40 +0.6*31=34.6
6 64 Step 2: A2 ” =0.4A2 + 0.6 A1” = 0.4*34.6 +0.6*31=32.44
7 58
Step 3: a2 =2A2 - A2” = 2*34.6 – 32.44 = 36.76
8 68
9 Step 4: b2 =0.4*(A2 - A2”)/0.6
= 0.4(34.6 – 32.44)/0.6 = 1.44
22
Step 5: F3 =a2 + b2 = 36.76 +1.44=38.2
 BROWN’S METHOD EXAMPLE
Forecast the demand for period 9 using Brown’s method (=0.4)

Week Actual demand Simple es Double es Forecast

t yt At At” at bt Ft
1 31 31.00 31.00 31.00 9.00
2 40 34.60 32.44 36.76 1.44 40.00
3 43 37.96 34.65 41.27 2.21 38.20
4 52 43.58 38.22 48.93 3.57 43.48
5 49 45.75 41.23 50.26 3.01 52.50
6 64 53.05 45.96 60.14 4.73 53.27
7 58 55.03 49.59 60.47 3.63 64.86
8 68 60.22 53.84 66.60 4.25 64.10
9 70.85
Brown’s Method

Period 2 – 8
24
SES: MAD = 10.43; MSE = 129.16
Brown: MAD = 5.47; MSE = 40.76
 HOLT’S METHOD

 Suitable for linear trend series

Yt = 0 + 1t + t
where 0 and 1 may change slowly with time.

 Calculation procedures for Holt’s method:

 Step 1: The current level of the data at time t, 0 , is estimated by
At  yt  (1  )( At 1  Tt 1 )
 Step 2: The slope of the series at time t, 1 , is estimated by
Tt  ( At  At 1 )  (1  )Tt 1

 Step 3: The forecast value for time period t+m made at time t is
Ft+m = At+ Tt m, m = 1, 2, 3,  25
where m is the forecast horizon.
 Choice of  ,  and Initialization in Holt's Method
 Choice of  and 
 Choose one that minimize MSE or MAPE.

 Initialization
 Delurgio
Level : A1 = y1
Trend :
 y2  y1    y4  y3 
T1 =
2

 Bowerman (P.403)
Fit a trend line to first few or one half of the historical data to find A0 and
T0.

26
 Holt's Method

 Advantage
•  is the smoothing constant for the level of the
series
•  is the smoothing constant for the trend
• apply different weights to actual data () and
trend ()

 Disadvantage
specify 2 parameters, not simple 27
 HOLT’S METHOD EXAMPLE

Forecast the demand for period 9 using Holt’s method (=0.5, =0.3)

Week Actual demand Smoothed value Smoothed trend Forecast

t yt At Tt Ft
1 31 31.00 9.00
2 40 t=1 40.0
3 43 Step 1: A1 = y1 =31
4 52
5 49 Step 2: T = { (y - y ) + (y - y ) } / 2
1 2 1 4 3
6 64 = { (40-31 ) + (52-43 ) } / 2 = 9
7 58
8 68
9 Step 3: F2 =A1 + T1 = 31 + 9=40 28
 HOLT’S METHOD EXAMPLE

Forecast the demand for period 9 using Holt’s method (=0.5, =0.3)

Week Actual demand Smoothed value Smoothed trend Forecast

t yt At Tt Ft
1 31 31.00 9.00
2 40 40.00 9.00 40.0
3 43 t=2 49.0
4 52 Step 1: A2 =0.5y2 + 0.5 (A1 + T1 )= 0.5*40 +0.5*(31+9)=40
5 49
6 64
Step 2: T2 =0.3(A2 - A1 ) + 0.7*T1 = 0.3*(40-31) +0.7*9=9
7 58
8 68
9 Step 3: F3 =A2 + T2 = 40 + 9=49 29
 HOLT’S METHOD EXAMPLE
Forecast the demand for period 9 using Holt’s method (=0.5, =0.3)

Week Actual demand Smoothed value Smoothed trend Forecast

t yt At Tt Ft
1 31 31.00 9.00
2 40 40.00 9.00 40.0
3 43 46.00 8.10 49.0
4 52 53.05 7.79 54.1
5 49 54.92 6.01 60.8
6 64 62.46 6.47 60.9
7 58 63.47 4.83 68.9
8 68 68.15 4.79 68.3
9 72.9

30
Holt’s Method

Period 2 – 8
SES: MAD = 10.43; MSE = 129.16 31
Brown: MAD = 5.47; MSE = 40.76
Holt : MAD = 4.89; MSE = 44.22
PART THREE
 Simple Exponential Smoothing
 Double Exponential Smoothing
 Brown’s One-parameter Linear Method
 Holt’s Two-parameter Linear Method
 Winters’ Three Parameter Linear and Seasonal
Exponential Smoothing
 Multiplicative Method
 Additive Method

32
 Winters' Multiplicative Method
• Suitable for linear trend and multiplicative seasonality
series 900
Winters' Multiplicative Method

800

700

Yt = (0 + 1t)*St + t 600

500

Actual
400

300

200

where 100

0
0 5 10 15 20 25 30
Period

0 , 1 , St may change slowly with time, and

the seasonal variation is increasing as the average
level of the series 0 + 1t increases. 33
WINTERS' MULTIPLICATIVE METHOD
 Calculation procedures for Winters' multiplicative method:
 Step 1: The current level of the data at time t, 0 , is estimated by
y
At  ( t )  1     At 1  Tt 1 
St  L
 Step 2: The slope of the series at time t, 1 , is estimated by
Tt    At  At 1   1    Tt 1

 Step 3: The seasonality of the data at time t, St , is estimated by

yt
St   ( )  1    St  L
At
 Step 4: The forecast value for time period t+m made at time t is
Ft  m   At  mTt  * St  m  L

Note: L is the length of seasonality (eg: L=4 for quarterly, L=12 for 34
monthly)
 Winters' Multiplicative Method

 Winters' exponential smoothing is an extension of Holt's linear

exponential smoothing. Winters' smoothing uses the 3
equations of Holt's model but introduces the seasonal index St
into the formulas and includes an extra equation that is used to
estimate seasonality.

 In step 1 equation, yt is divided by St–L . This removes the

seasonal effects which may exist in the original data yt.

 Step 2 equation is just the same as estimate of trend equation

in Holt’s method. 35
 Winters' Multiplicative Method
 The form of step 3 equation is similar to that of all other
exponential smoothing equations i.e. a value  in this case yt /At 
is multiplied by a constant  and is then added to its previous
smoothed estimate which has been multiplied by 1  . Note that
the estimate of seasonality is given as an index, fluctuating around
1.

 Step 4 equation is similar to the forecast equation in Holt’s

method except that the estimate for the future period, t + m, is
multiplied by StL +m . This is the last seasonal index available and
hence is used to readjust the forecast for seasonality. Multiplying
the forecast by St L +m has the opposite effect of dividing yt by StL
36
in step 1 equation.
Choice of , ,  in Winters' Multiplicative Method

  is the smoothing constant for the level of the

series
  is the smoothing constant for the trend
  is the smoothing constant for the seasonality.

 Choose , ,  which minimize MSE or MAPE.

37
Initialization in Winters' Multiplicative Method

 Makridakis (P.168)
1
AL   y1  y2  ...  y L 
Level : L

Trend : 1  y L1  y1 y L  2  y2 y L L  y L 
TL     ...  
L L L L
Seasonal :
y1 y2 yL
S1  , S2  , ... , S L 
AL AL AL
 Bowerman (P.403-7)
Yr/Season 1 2 … L

1 y1 y2 … yL
38
2 yL+1 yL+2 … y2L
Winters’ Multiplicative Example : Quarterly Exports of a French Company( = 0.822,  = 0.055 and  = 0.05)

Period Actual Level Trend Seasonal Forecast Winters' Multiplicative Method

1 362 – – 0.953 – 900

2 385 – – 1.013 – 800

700
3 432 – – 1.137 – 600

4 341 380.00 9.75 0.897 – 500

Actual
5 382 371.29 400

6 409 300

7 498
t= 4 200

100

8 387 Step 1: A4 = (y1 + y2 +y3 + y4)/4 =(362+385+432+341)/4= 380 0

0 5 10 15 20 25 30
9 473 Period

10 513
11 582 Step 2: T4 =[(y5 - y1)/4 + (y6 – y2)/4+(y7 - y3)/4 + (y8 – y4)/4]/4
12 474 = [(382-362) + (409-385)+ (498-432)+ (387-341)]/16
13 544
14 582 = 9.75 Yr/ 1 2 3 4
15 681 Qtr
16 557 1 362 385 432 341
17 628 Step 3: S1 = y1 /A4 = 362/380 = 0.953
2 382 409 498 387
18 707 S2 = y2 /A4 = 385/380 = 1.013
19 773
20 592
S3 = y3 /A4 = 432/380 = 1.137
21 627 S4 = y4 /A4 = 341/380 = 0.897
22 725
23 854
24 661 Step 4: F5 =(A4 + T4)* S1 = (380 + 9.75)*0.953=371.29
25
Winters’ Multiplicative Example : Quarterly Exports of a French Company( = 0.822,  = 0.055 and  = 0.05)
Perio Actua Level Trend Seasonal Forecast Winters' Multiplicative Method

d l 900

1 362 – – 0.953 – 800

700
2 385 – – 1.013 – 600

3 432 – – 1.137 – 500

Actual
4 341 380.00 9.75 0.897 – 400

300
5 382 398.99 10.26 0.953 371.29 200

6 409 414.64 100

7 498 t=5 0
0 5 10 15 20 25 30

8 387 Period

9 473 Step 1: A5 = 0.822y5 / S1 + 0.178 (A4 + T4 )

10 513 = 0.822*382/ 0.953+0.178*(380+9.75)=398.99
11 582
12 474
13 544
Step 2: T5 =0.055(A5 - A4 ) + 0.945*T4
14 582
15 681 = 0.055*(398.99-380) +0.945*9.75=10.26
16 557
17 628
18 707 Step 3: S5 =0.05y5 / A5 + 0.95*S1
19 773
20 592 = 0.05*(382/398.99) + 0.95*0.953=0.953
21 627
22 725
23 854 Step 4: F6 = (A5 + T5 )* S2 = (398.99+10.26)*1.013=414.64
24 661
25
Winters’ Multiplicative Example : Quarterly Exports of a French Company( = 0.822,  = 0.055 and  = 0.05)
Period Actual Level Trend Seasonal Forecast
1 362 – – 0.953 –
2 385 – – 1.013 –
3 432 – – 1.137 –
4 341 380.00 9.75 0.897 –
5 382 398.99 10.26 0.953 371.29
6 409 404.68 10.01 1.013 414.64
7 498 433.90 11.06 1.137 471.43
8 387 433.70 10.44 0.897 399.29
9 473 487.09 12.81 0.953 423.21
10 513 505.24 13.10 1.013 506.42
11 582 512.88 12.80 1.137 589.56
12 474 527.88 12.92 0.897 471.60
13 544 565.10 14.26 0.954 515.81
14 582 575.32 14.04 1.013 586.98
15 681 597.13 14.46 1.137 670.25
16 557 619.20 14.88 0.897 548.69
17 628 653.85 15.97 0.954 605.06
18 707 692.88 17.24 1.013 678.57
19 773 685.04 15.86 1.137 807.70
20 592 667.10 14.00 0.897 628.90
21 627 661.18 12.90 0.954 650.13
22 725 708.03 14.77 1.014 683.14
23 854 746.08 16.05 1.137 821.80
24 661 741.54 14.92 0.897 683.47
25 721.83 (m = 1)
 Winters' Additive Method
 Suitable for linear trend and additive seasonality series

yt = (0 + 1t) + St + t
Winters' Additive Method

60

50

40

Actual
30

20

10

where
0
0 5 10 15 20 25
Period

0 , 1 , St may change slowly with time and the seasonal

variation is constant over time.
42
WINTERS' ADDITIVE METHOD
 Calculation procedures for Winters' additive method:
 Step 1: The current level of the data at time t, 0 , is estimated by

At    yt  St  L   1     At 1  Tt 1 
 Step 2: The slope of the series at time t, 1 , is estimated by
Tt    At  At 1   1    Tt 1
 Step 3: The seasonality of the data at time t, St , is estimated by
St    yt  At   1    St  L

 Step 4: The forecast value for time period t+m made at time t is
Ft  m  ( At  mTt )  St  L m 43

Note: L is the length of seasonlity

Choice of , ,  and Initialization in Winters’
Additive Method

 Makridakis (P.169)
Same as those for multiplicative method except for seasonal
indices, we use
S1 = y1 – AL , S2 = y2 – AL , … , SL = yL – AL

 Bowerman (P.417)
Compute the least squares estimates of the parameters in the
dummy variable regression model
y t   0   1t   S 1 x S 1,t   S 2 x S 2 ,t     S ( L 1) x S ( L 1),t   t
where independent error terms are assumed.

44
Winters’ Additive Example : Product Demand ( = 0.2,  = 0.1 and  = 0.1)
Winters' Additive Method
t yt At Tt St(t) Forecast 60

1 10 -- -- -15 -- 50

2 31 -- -- 6 -- 40

3 43 -- -- 18 --

Actual
30

4 16 25.00 0.375 -9 -- 20

10

5 11 10.375 0

6 33 t=4
0 5 10
Period
15 20 25

7 45 Step 1: A4 = (y1 + y2 +y3 + y4)/4 =(10+31+43+16)/4= 25

8 17
9 13
10 34 Step 2: T4 =[(y5 - y1)/4 + (y6 – y2)/4+(y7 - y3)/4 + (y8 – y4)/4]/4
11 48 = [(11-10) + (33-31)+ (45-43)+ (17-16)]/16= 0.375
12 19
13 15 Step 3: S1 = y1 - A4 = 10 - 25 = -15
14 37 Yr/ 1 2 3 4
15 51 S2 = y2 - A4 = 31 – 25 = 6 Qtr

16 21 S3 = y3 - A4 = 43 – 25 = 18 1 10 31 43 16
2 11 33 45 17
17 16 S4 = y4 - A4 = 16 – 25 = -9
18 39
19 53 Step 4: F5 =(A4 + T4) +S1 = (25 + 0.375) - 15=10.375
20 22
Winters’ Additive Example : Product Demand ( = 0.2,  = 0.1 and  = 0.1)
t yt At Tt St(t) Forecast Winters' Additive Method

1 10 -- -- -15 -- 60

50
2 31 -- -- 6 -- 40

3 43 -- -- 18 --

Actual
30

4 16 25.00 0.375 -9 -- 20

5 11 25.50 0.388 -14.950 10.375 10

6 33 31.888 0 5 10
Period
15 20 25

7 45 t=5
8 17
Step 1: A5=0.2(y5 - S1)+ 0.8 (A4 + T4 )
9 13
10 34 = 0.2*[11-(-15)] + 0.8 * (25+0.375) = 25.5
11 48
12 19 Step 2: T5 =0.1(A5 - A4 ) + 0.9*T4
13 15 = 0.1*(25.5-25) +0.9*0.375 = 0.388
14 37
15 51
16 21 Step 3: S5 =0.1(y5 - A5 )+ 0.9*S1
17 16 = 0.1*(11-25.5) + 0.9*(-15) = -14.95
18 39
19 53 Step 4: F6 = (A5 + T5 )+ S2 = (25.5+0.388)+ 6 = 31.888
20 22
Winters’ Additive Example : Product Demand ( = 0.2,  = 0.1 and  = 0.1)
t yt At Tt St(t) Forecast
1 10 -- -- -15 --
2 31 -- -- 6 --
3 43 -- -- 18 --
4 16 25.00 0.375 -9 --
5 11 25.50 0.388 -14.950 10.375
6 33 26.11 0.410 6.089 31.888
7 45 26.62 0.419 18.038 44.520
8 17 26.83 0.399 -9.083 18.035
9 13 27.37 0.413 -14.892 12.277
10 34 27.81 0.416 6.099 33.874
11 48 28.57 0.450 18.177 46.264
12 19 28.84 0.432 -9.158 19.940
13 15 29.39 0.444 -14.842 14.374
14 37 30.05 0.465 6.184 35.935
15 51 30.98 0.512 18.362 48.692
16 21 31.22 0.485 -9.264 22.329
17 16 31.53 0.468 -14.911 16.864
18 39 32.16 0.484 6.249 38.186
19 53 33.05 0.524 18.521 51.010
20 22 33.11 0.478 -9.449 24.305
21 18.675
Period 5 – 20
Additive : MAD = 1.13; MSE = 1.67
Multiplicative : MAD = 1.86; MSE = 5.81 48